[ARC345]

Summary

• Adds a Smart Context plugin that extracts topics from chat conversations via LLM and displays them as a clickable bubble cloud in a side panel
• Clicking topic bubbles activates an embedding-based filter that limits which messages are injected into the LLM pipeline as history (threshold-based cosine similarity)
• "Instructions & Preferences" messages (short directives like "be concise") are identified by the LLM and always injected regardless of filter state
• Chat messages matching the active filter get a blue left-border highlight in the main chat view

New files

File	Purpose
example-vault/plugins/smart_context/main.py	Full plugin: pipeline subscribers, commands, embedding cache usage
example-vault/plugins/smart_context/embedding_cache.py	File-backed JSON embedding cache keyed by {msg_id}:{content_hash}
example-vault/plugins/smart_context/panel.html	Bubble cloud UI with similarity toggle and N-of-M status line

Modified files

File	Change
sidecar/services/llm_service.py	Added embed() method via LiteLLM aembedding
src/vault/chat/chat.js	Added setupSmartContextEventListeners() for message highlighting
src/styles/styles.css	Added .sc-highlight class
example-vault/.vault.toml	Added explicit smart_context config defaults

Test plan

• 158 tests pass (pixi run test)
• Open example-vault, send a few messages on different topics
• Smart Context panel appears on right side with topic bubbles after first response
• Clicking a bubble highlights matching messages in chat and filters LLM context
• "Instructions & Preferences" bar shows count of always-on messages
• Similarity toggle disables filtering (bubbles become visualization-only)
• Status line shows "N of M messages in context" when filter active

🤖 Generated with Claude Code